UAV Multispectral Imagery Can Complement Satellite Data for Monitoring Forest Health - MDPI
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
remote sensing
Article
UAV Multispectral Imagery Can Complement
Satellite Data for Monitoring Forest Health
Jonathan P. Dash 1,∗ , Grant D. Pearse 1 and Michael S. Watt 2
1 Scion, 49 Sala Street, Private Bag 3020, Rotorua 3046, New Zealand; grant.pearse@scionresearch.com
2 Scion, 10 Kyle St, P. O. Box 29237, Christchurch 8440, New Zealand; michael.watt@scionresearch.com
* Correspondence: jonathan.dash@scionresearch.com
Received: 17 June 2018; Accepted: 1 August 2018; Published: 3 August 2018
Abstract: The development of methods that can accurately detect physiological stress in forest trees
caused by biotic or abiotic factors is vital for ensuring productive forest systems that can meet
the demands of the Earth’s population. The emergence of new sensors and platforms presents
opportunities to augment traditional practices by combining remotely-sensed data products to
provide enhanced information on forest condition. We tested the sensitivity of multispectral
imagery collected from time-series unmanned aerial vehicle (UAV) and satellite imagery to detect
herbicide-induced stress in a carefully controlled experiment carried out in a mature Pinus radiata
D. Don plantation. The results revealed that both data sources were sensitive to physiological stress
in the study trees. The UAV data were more sensitive to changes at a finer spatial resolution and
could detect stress down to the level of individual trees. The satellite data tested could only detect
physiological stress in clusters of four or more trees. Resampling the UAV imagery to the same
spatial resolution as the satellite imagery revealed that the differences in sensitivity were not solely
the result of spatial resolution. Instead, vegetation indices suited to the sensor characteristics of
each platform were required to optimise the detection of physiological stress from each data source.
Our results define both the spatial detection threshold and the optimum vegetation indices required
to implement monitoring of this forest type. A comparison between time-series datasets of different
spectral indices showed that the two sensors are compatible and can be used to deliver an enhanced
method for monitoring physiological stress in forest trees at various scales. We found that the higher
resolution UAV imagery was more sensitive to fine-scale instances of herbicide induced physiological
stress than the RapidEye imagery. Although less sensitive to smaller phenomena the satellite imagery
was found to be very useful for observing trends in physiological stress over larger areas.
Keywords: tree health; precision forestry; sensor fusion; RPAS; drone; RapidEye; plantation forest;
radiata pine; forest management; forest productivity
1. Introduction
Highly productive plantation forests are required to meet the timber and fibre demands
of the Earth’s population in a sustainable manner. A significant research effort is focussed on
maximising the returns from global forest plantations through silvicultural practices [1,2], tree breeding
programmes [3,4], improving forest nutrition [5], and developing remote sensing methods to enhance
forest management [6–8]. Yields from intensively managed monoculture plantations face significant
threats from unwanted organisms, the impact of extreme weather events, and increasingly from the
adverse effects of air pollution [9,10]. The frequency and severity of losses associated with these factors
are expected to increase in many regions as the effects of climate change become more severe [11].
Tools to detect, quantify, and help mitigate the impacts of these damaging biotic and abiotic factors
are essential for maximising plantation forest yields. In recent years, significant progress has been
Remote Sens. 2018, 10, 1216; doi:10.3390/rs10081216 www.mdpi.com/journal/remotesensing
Remote Sens. 2018, 10, 1216 2 of 22
made towards the detection and monitoring of pests and diseases in trees and forests. These methods
often use remote sensing and ground assessments to detect signs and symptoms of a stressor [12].
Surveillance procedures are required for a range of spatial scales to detect incursions and to inform
effective forest management to mitigate negative effects on tree growth and the probability of tree
mortality [13,14].
Since the advent of civilian Earth-observing satellites suitable for vegetation monitoring,
a significant body of research has used satellite imagery to quantify the spatial extent of insect outbreaks
and tree mortality at regional and landscape scales using moderate (5–30 m) and coarse (≥30 m)
resolution imagery [15–18]. More recently, research has established the utility of high-resolution (≤5 m)
satellite imagery [19–21] and imagery collected from manned aerial platforms [14,22,23] for monitoring
tree health. This research has shown that sensors with finer spatial resolution offer greater accuracy for
detecting and monitoring disease expression. This is especially true for early-stage outbreaks or those
with a limited spatial extent that may only be apparent at the sub-stand or individual tree level.
Unmanned aerial vehicles (UAVs) have emerged as a useful data source applicable to many forest
management scenarios [24]. A growing number of studies have examined the use of UAVs for a wide
range of applications including forest measurement [25–30], informing silvicultural practice [31,32],
forest health monitoring [12,33–35], and monitoring events such as fires [36,37], wind damage [38],
and harvesting [39]. The cost of these platforms and associated sensors are rapidly decreasing while
their capabilities and sophistication are constantly improving. There are several appealing aspects of
UAVs including greater flexibility due to ease of deployment and transportation, ability to capture
extremely high resolution data, and the capacity to capture useful imagery even in the presence of cloud
cover. UAVs cannot rival traditional platforms in terms of spatial extent; however, with appropriate
processing and analysis frameworks, data from UAVs are likely to offer substantial opportunities to
augment and enhance data collected from these more traditional platforms.
A few studies have sought to develop methods linking UAV data and satellite imagery to provide
an enhanced data product with utility beyond that of the component datasets. Imagery collected from
UAVs has been used to calibrate measures of burn severity derived from Landsat 8 imagery following
a wildfire with some success [40]. UAV imagery has also been used to augment satellite imagery
to inform land managers about compliance with resource management policy in riparian habitats.
The UAV imagery enhanced both the temporal and spatial resolution of the satellite imagery providing
an effective tool for monitoring compliance and for conservation of endangered habitat [41]. A study
based in Sumatra [42] used a combination of Landsat and UAV imagery to monitor orangutan habitat.
The UAV data were used to train classification algorithms applied to Landsat data to differentiate
land use classes (oil palm plantation vs. reforestation vs. logged forest) relevant to orangutan
conservation with considerable success (classification accuracy over 75%) [42]. In a similar manner,
a study conducted in a Costa Rican ecological reserve used UAV data to identify land cover classes
and then applied these to a historical archive of Landsat data [43]. Methods have been proposed for
integrating imagery from the European Space Agency’s Sentinel-2 satellites with UAV data for forest
inventory purposes. In the aforementioned study, a hierarchical, model-based mode of inference was
used to provide viable population estimates and corresponding uncertainties based on datasets with
differing spatial coverages [44]. Initial methods have also been developed to extrapolate UAV data on
tree measurements using imagery from the Pléiades satellite constellation [45].
Change detection techniques provide a valuable starting point for the identification of subtle
changes in forest health based on the spectral information. The aggregation of multi-temporal images
can greatly improve the signal-to-noise ratio [46] and the use of longer time periods can mitigate
the confounding impacts of environmental factors such as cloud cover on detection. Using this
technique, acceptable accuracies have been demonstrated for the detection of bark beetle attack
on conifer species [47–50] and needle discolouration resulting from Dothistroma pini [14]. Analysis
of time-series satellite imagery was used by Eitel et al. (2011) who chainsaw girdled patches of
piñon-juniper woodland and then successfully monitored the induced tree stress before mortality
Remote Sens. 2018, 10, 1216 3 of 22
using a dense sequence of multispectral images from the RapidEye (Planet, San Francisco, USA) satellite
constellation [51]. Invoking physiological changes in this manner provides superior experimental
control and is well suited to testing and calibrating methods for the early detection of symptoms such
as pathogen-induced physiological changes in foliage. In a similar fashion, Dash et al. [12] injected
herbicide into the stems of mature plantation trees and tracked changes in canopy colour and density
using time-series multispectral UAV imagery. This study [12] confirmed the utility of high-resolution,
time-series UAV data for monitoring physiological stress in trees and identified suitable spectral indices
for this task. However, these findings were limited to small areas due to the limited data collection
capability of UAV platforms. Further research was required to investigate whether widely-used
satellite imagery products could be used as a complementary data source to expand the detection of
physiological stress to larger areas.
In this study, we investigated the utility of UAV and satellite multispectral imagery for detecting
physiological stress in forest trees and sought to define the detection capability. To accomplish this,
we developed methods to link UAV and satellite data for detection of physiological stress and tested
whether these could provide increased area coverage and models with greater predictive power
than either dataset alone. We hypothesised that very high resolution UAV imagery collected with a
narrow-band multispectral camera could provide greater predictive capability for detecting tree stress
than broad-band satellite imagery. We also tested whether differences in sensitivity were associated
with differences in the spatial or spectral resolution of the sensors.
2. Methods
The study site, experimental treatment, and field data were the same as reported in [12].
The essential details are reproduced here for convenience.
2.1. Study Site
As described in [12], the study site was located in Kinleith Forest in New Zealand’s Central
North Island (Figure 1) (Latitude 38◦ 240 18.74S, Longitude 176◦ 00 59.28E), approximately 28 km
south-east of the town of Tokoroa. The site elevation is approximately 230 m and slopes gradually up
towards the south-eastern corner of the trial. The site experiences a temperate maritime climate
(total annual rainfall = 1238 mm, mean annual temperature = 13.4 ◦ C) and the soils are loam
belonging to the Kairanga series [52]. The trial area of interest was 2.7 ha and was located within
a 7.3 ha stand of Pinus radiata D. Don (P. radiata) planted in July 1993. Site preparation was via
V-blading and the site was in its second rotation under plantation forest. Stand density at planting
was 635 stems/ha (4.5 m × 3.5 m). No thinning was undertaken and trees that suffered mortality
early in the rotation were replaced. The site contained 35 rectangular (0.04 ha) plots that were part
of a now defunct fungicide spray trial into disease resistance in immature trees. The fungicide
spray trial was completed before age 6 (1999) whereafter the trees were left to grow. At the
initiation of this study, a comprehensive health survey of the site showed no evidence of disease
expression and the majority of trees were in good health. The study trees were mature and large
(mean diameter at breast height = 462 mm, mean top height = 43.5 m) at the time of data collection
and were harvested soon after the conclusion of this research.
2.2. Experimental Design
The existing trial layout and plots were used to distribute the experimental treatments throughout
the study area. These were used because they were clearly demarcated and accurately mapped, making
design and navigation simple. Of the 35 plots five plots were randomly assigned to each of 6 treatments
(Table 1), with the remaining 5 plots were excluded from the trial. All plots were rectangular and
contained 35 trees in seven rows planted with a spacing of 3.5 × 4.5 m. Trees that were dead or severely
suppressed were not included in the trial. The mean number of dead, severely suppressed, or missing
trees per plot was 5.79 (range = 2–11). The experimental treatment defined the size of the cluster
Remote Sens. 2018, 10, 1216 4 of 22
(groups of treated trees) that were poisoned to cause physiological stress and simulate the symptoms
that might be expected during a disease outbreak. The objective of using different sized clusters was
to investigate the sensitivity of different remotely sensed datasets to different levels of physiological
stress as might be associated with certain diseases. Study trees were treated with Metsulfuron methyl
in the form of water-dispersible granules at a rate of 200 g/L. This mixture was injected through holes
drilled into the tree stem. Holes were drilled 25 cm apart around the circumference of each stem
and 15 mL of herbicide was deposited into each hole. In total, 13.32 L of Metsulfuron mixture was
applied to the study site. Administration by injection was used as it allowed for carefully targeted
application and because expert advice indicated it would invoke changes in foliar colour and retention
within the duration of the study. Herbicide application was completed for all trees in a single day on
the 3 December 2015 [12].
Figure 1. Location of the study site within New Zealand with examples of the broadband multispectral
imagery from RapidEye (a) and the UAV MicaSense Red Edge camera (c) with both images showing
the trial area at a 1:5000 scale. The normalised spectral response curves for the RapidEye (b) and the
MicaSense RedEdge (d) sensors are also shown based on data provided by the sensor developers.
Table 1. Experimental treatments used in the study. The number of trees subjected to herbicide
application in the treatment and the mean cross-sectional area occupied by the canopy of the treated
trees is shown. This table is modified from [12].
Treatment Description n Area (m2 )
0 Control, no trees poisoned 0 575
1 Single tree closest to plot centre is poisoned 1 18
2 Two trees closest to the plot centre poisoned 2 30
3 Four trees closest to the plot centre poisoned 4 77
4 Eight trees closest to the plot centre poisoned 8 126
5 Sixteen trees closest to the plot centre poisoned 16 279
Remote Sens. 2018, 10, 1216 5 of 22
2.3. UAV Imagery
The UAV dataset used in this study was originally described in [12]. A co-axial quad-copter UAV
was used for all data collection. The craft was a modified version of the Aeronavics (Aeronavics, Raglan,
New Zealand www.aeronavics.co.nz) SkyJib heavy-lift airframe. This craft had a flying time of around
15 min. with a payload of approximately 5 kg and provided a stable platform even in strong, turbulent
winds. A single flight plan was used for all data collections to minimise variance associated with
campaign settings. The flight plan used 18 flight lines to cover the area of interest and ensured that all
craft manoeuvres and altitude adjustments occurred outside of the study area to minimise flight-related
artefacts in the data collected. The original intent was to collect imagery on a weekly basis as the
weather conditions permitted. In the end, data collection was significantly sparser due to technical
difficulties and poor weather resulting in nine separate acquisitions between 30 October 2015 and
3 March 2016.
A series of nine ground control points (GCPs) were established around the study area and fixed
using a Trimble Geo7X GNSS receiver (Trimble Navigation Ltd., Sunnyvale, CA, USA). Positions were
differentially corrected using a network of local continuously operating reference stations maintained
by Land Information New Zealand (LINZ). The targets established at the ground control points were
easily manually identified in the multispectral imagery. The GCPs were used to accurately georectify
all UAV imagery [12].
Narrowband multispectral imagery was collected using a MicaSense RedEdge 3 camera
(Micasense, WA, USA). The camera provides imagery in five narrow bands (blue = 455–495 nm,
green = 540–580 nm, red = 658–678 nm, red edge = 707–727 nm and near-infrared = 800–880 nm)
(Figure 1) via five separate imaging units that operate nearly simultaneously. The camera has a lens
focal length of 5.5 mm and field of view of 87.4◦ . The camera was housed in a gimbal to ensure nadir
orientation of the camera during data collection. The flight plan ensured cross-track and along-track
overlap of 85 % and a calibrated reflectance panel was imaged directly before and after each flight and
used for reflectance calibration using the empirical line method [53]. The UAV imagery was collected
from a flying altitude of approximately 90 m above the local terrain and resulted in a ground sample
distance (GSD) of 6 cm. Imagery acquisition was limited to within two hours either side of local solar
noon to minimise shadows.
The imagery was exported from the MicaSense RedEdge 3 in 12-bit RAW format. These images
were mosaicked into a single, multi-band image covering the entire study area. The raw digital numbers
(DNs) were converted into reflectance values using the calibrated reflectance panel. The relationship
between DNs and the natural logarithm of the image surface reflectance is linear and the y-intercept
can be interpreted as the minimum surface reflectance that can be detected for each wavelength
band [53]. This can then be used to develop a regression equation to convert DNs to reflectance values
using a calibration equation [54]. Images were georectified using the GCPs recorded within the study
site and visible in the imagery [12].
2.4. Satellite Imagery
RapidEye is a constellation of Earth-observing satellites that provide sun synchronous imagery
from an orbit height of 680 km with a swath width of 77 km and a 5 day revisit time. RapidEye imagery
has a spatial resolution of 5 m and provides data in 5 broad bands (blue = 440–510 nm,
green = 520–590 nm, red = 630–685 nm, red edge = 690–730, and near-infrared 760–850 nm) (Figure 1).
Four RapidEye (Planet, San Francisco, CA, USA) satellite images that included the study site were
acquired from the 28 November 2015, 10 December 2015, 20 December 2015, and 20 February 2016.
These were the only suitable cloud-free images available during the duration of the experiment. Level
3A RapidEye images were used as an input with a pixel size of 5.0 m. The images were radiometrically
and geometrically corrected by the supplier. A noise filter and top-of-atmosphere (TOA) dark object
subtraction were applied to each image. The residual geographic error was corrected by manually
geo-registering all images to a reference UAV image using a polynomial warping method (ENVI 4.5,
Remote Sens. 2018, 10, 1216 6 of 22
Harris, Boulder, CO, USA). The total RMS Error after applying the correction across the entire scene
was 0.29 m.
2.5. Image Processing
Several spectral indices were chosen and calculated from each set of images following a review of
the literature on the topic of stress detection using multispectral imagery (Table 2). Spectral indices were
selected for inclusion based on the spectral bands available in both datasets and evidence of correlation
with plant physiological stress from previous research. The normalised difference vegetation index
(NDVI) is a very well known and widely used ratio that has been shown many times to be correlated
with plant condition, physiological stress, and photosynthetic activity in a wide range of vegetated
environments [55–59]. The green normalised difference vegetation index (GNDVI) and the red edge
normalised difference vegetation index (RENDVI) were included specifically to examine the sensitivity
of indices including the red edge and green bands from each sensor and to compare these directly to
an index based on the red band (NDVI).
Table 2. Spectral indices calculated from the multispectral imagery. This table is modified from [12].
Spectral Index Equation Source
( N IR− RED )
NDVI ( N IR+ RED )
[60]
( N IR− GREEN )
GNDVI N IR+ GREEN [61]
( N IR− REDEDGE)
RENDVI N IR+ REDEDGE [62,63]
As data collection spanned several months, methods were required to account for background
changes in the target vegetation that were not caused by the experimental treatments. These changes
include physiological response to the growing season, changes caused by climatic events,
and atmospheric effects that could not be controlled for using other methods [12]. Without a suitable
method for control, these influences would confound the physiological stress signal induced by the
application of herbicide. Following the method of Healey et al. [64], spectral indices were de-trended
to remove the extraneous effects. The approach was based on the assumption that changes in adjacent
but untreated parts of the study site could be assumed to be caused by factors unrelated to the
herbicide-induced stress. By characterising these unrelated trends we could minimise the potential
impact on any signal from the stressed study trees [51].
For the UAV imagery, spectral indices were de-trended by identifying 25 regions of the study
area in the final orthomosaic to act as de-trending control areas. A review of the full set of time-series
mosaics was carried out to carefully select areas that covered conditions representative of the study
plots and the trends in the multi-temporal imagery. The control areas were 5 m2 and were carefully
placed to ensure these areas did not overlap the experimental tree clusters. For the RapidEye data,
control areas were selected from a nearby P. radiata stand with similar growing conditions, terrain,
and tree age. Thirty control areas were carefully selected and used to de-trend the RapidEye imagery.
For both sets of imagery, the spectral indices extracted from the control areas were deemed to represent
the average conditions of the study forest within a specific time step. Spectral indices extracted from
both the treated and untreated plots within the experiment were rescaled to the standard deviation
above or below the mean forest value for the spectral indices as extracted from the de-trending control
areas [64] as follows:
Si = (S − Sµ )/Sδ (1)
where Si is the rescaled spectral index, S is the original spectral index value, Sµ is the mean value
for the control areas within a time step and Sδ is the standard deviation of values from the control
Remote Sens. 2018, 10, 1216 7 of 22
areas. Using this method, the time-series spectral indices can be de-trended in a robust and easy to
implement manner provided that the control areas selected are representative of the average forest
conditions within a time step [64].
Raster layers representing the three spectral indices (Table 2) were calculated from each UAV
image mosaic and from each satellite image. These were clipped to the boundaries of the study
tree canopies that had been carefully delineated using the final image in each time series through
digitisation in a GIS (ESRI, Redlands, CA, USA). In the final image, treated trees could be easily and
precisely identified. The UAV imagery and the satellite imagery were digitised separately to account
for some residual variation caused by differences in sensor geometry. The values for all pixels within
the delineated boundaries for treated trees were averaged to provide a mean for each cluster at each
acquisition date. For the control (untreated) plots pixels were averaged within the boundary of the
original 35 tree plot.
The spectral indices from both datasets were de-trended using the method described earlier [64].
Analysis of the UAV imagery indicated that there was a significant trend in the data that was not due
to the herbicide-induced stress [12]. Once the background changes were accounted for, the trend in
spectral indices in the treated tree clusters was more pronounced, suggesting that the background
noise masking the spectral stress signal had been adequately controlled.
Changes in spectral indices were defined using an approach developed in previous
experiments [12,51]. This approach defined a “no-change” zone between the 25th and the 75th
percentile of a de-trended spectral index for a given observation at the start of the time-series.
Subsequent observations of the same spectral index can be deemed to be stressed if the 75th percentile
of the spectral index values for a given observation drops below the lower bound of the “no-change”
region defined at the start of the time-series.
To compare the influence of spectral and spatial resolution directly on the detection of
physiological stress the UAV imagery was resampled to the same spatial resolution as the RapidEye
data (5 m). A cubic spline resampling was applied to all UAV images using the gdalwarp function of
the geospatial data abstraction library (GDAL version 2.2.1) [65] using the standalone version of the
GDAL software. The resulting images were manually checked for any visual artefacts associated with
the resampling process. All imagery was found to be artefact free and spectral indices were calculated
and extracted using the same processing chain applied to the original UAV imagery.
2.6. Tree Health Data
The method for field-based, manual tree health monitoring was first described in detail in [12].
Tree canopies of all study trees were assessed for needle discolouration and canopy density from the
ground by an experienced assessor who is regularly engaged in forest health monitoring. The assessor
undergoes regular training, auditing, and calibration and carries photographic reference material to
improve consistency. To reduce the variance associated with assessment subjectivity the same expert
assessor was used for all assessments for the duration of the trial. Each tree was assessed from two
separate angles, approximately 180◦ from each other, so that a comprehensive assessment of the tree
canopy could be made. Assessment locations were permanently marked and trees were assessed from
these positions at approximately the same time of day at each time. The field assessor followed a
commonly used approach to classifying stress symptom expression by scoring percentage canopy
discolouration and density into increments of 5. Wherever possible, ground assessments were made on
the same day that UAV data were collected, the longest gap between ground and UAV data collection
was three days.
2.7. Statistical Analysis
Due to the revisit time of the RapidEye satellites and significant cloud cover in the study area,
there were several instances where satellite imagery was not available for days where UAV and ground
data had been acquired. As a result, the experimental dataset can be considered as a multivariate timeRemote Sens. 2018, 10, 1216 8 of 22
series dataset that contains missing values for each observer on certain dates. Analysis of datasets
with these characteristics is well studied with considerable research generated from diverse fields
such as econometrics [66] and hydrology [67]. Numerous approaches have been developed to estimate
the missing time series values to provide a complete dataset for subsequent analysis. All of these
methods rely on developing statistical models between variables in the time series data and an ability
to model the longitudinal component of the dataset through time. Methods to estimate missing
values include multivariate imputation via chained equations [68], regression-based methods [69],
and non-parametric, ensemble methods such as Random Forests [70]. Following an examination of
several available methods, imputation based on the Random Forest algorithm via the MissForest R
package [71] was selected for use in this study. A complete time series data set was produced with an
observation of ground, UAV, and satellite imagery at every acquisition date. This enabled the time
series datasets to be plotted through time and a direct exploration of the relationships between all
three datasets to be made.
The relationship between UAV and satellite derived spectral indices was explored using regression
models. The purpose of the models was to understand the relationship between the two different
sensor platforms and to develop predictive models that can be useful for extrapolating information
on tree stress derived from the UAV data to larger areas. Ordinary least squares (OLS) was used
to produce an initial model. Initial model diagnostic plots were carefully checked for evidence of
non-compliance with the assumptions of linear regression analysis and a Durbin-Watson test was used
to check for autocorrelation in the model residuals. No evidence of non-compliance was found and so
model fitting proceeded using OLS. Univariate models were developed with UAV spectral index as
the response and satellite spectral index as the predictor variable. Models up to and including the fifth
order polynomial were fitted and the best model for each spectral index was selected based on the
Akaike information criterion (AIC) outputted from the R statistical software (version 3.4.2) [72].
Random Forest (RF) regression models [70] were used to examine the capacity of spectral
indices extracted from both platforms to detect stress symptoms (canopy discolouration) measured
in the field. RF regression utilises ensemble decision tree classifiers, based on bootstrap aggregated
sampling (bagging), to construct many individual decision trees, from which a final class assignment is
determined [70]. RF has previously been used to successfully model several plantation forest variables
using remotely sensed data [6,7,73,74]. The RF algorithm constructs decision trees using a bootstrap
sample from the available training data, with the remaining data assigned as out-of-bag (OOB) samples.
At each node, a random subset of predictor variables is tested to partition the observation data into
increasingly homogeneous subsets. The node-splitting variable selected from the variable subset
is that which resulted in the greatest increase in data purity (variance or Gini) before and after the
tree node split [75]. The OOB sample data are used to compute accuracies and error rates averaged
over all predictions, and estimate variable importance [75,76]. RF provides methods to estimate
the importance of each predictor variable in the model. The mean decrease in accuracy (MDA)
importance measure is calculated as the normalised difference between OOB accuracy of the original
observations to randomly permuted variables [75–77]. RF is a well-regarded machine learning tool
that has the capacity to identify complex and non-linear relationships in the fitting dataset and offers
high classification accuracy [75,77,78].
RF models were fitted using the randomForest R package [79], the initial model included all
three spectral indices from both platforms and aggregated all data from the entire time-series dataset.
The model performance and the relative importance of the predictors were used to provide insight
into the sensitivity of the spectral indices to physiological stress. In a subsequent iterative processing
step, RF models were fitted using only UAV predictors and only satellite-derived predictors to assess
the relative utility of these data. The mean square error calculated from model out-of-bag errors and
the coefficient of determination (R2 ) were used to assess model precision. Conditional random forest
variable importance calculated using the party R package [80–82] were used to provide insight into
the predictive power of metrics from data from both sources. Conditional importance scores wereRemote Sens. 2018, 10, 1216 9 of 22
used as the variable importance scores calculated from RF have been shown to be biased in some
situations [81].
3. Results
3.1. Data Summary
In total, 25 separate datasets were collected, processed, and available for inclusion in this analysis.
In most cases, the field data and the UAV multispectral imagery were recorded on the same date.
All UAV imagery from the 27 November was abandoned as unfavourable weather conditions made
the data unusable. After March 2016, the study trees were harvested and the study was concluded as
no further data could be collected. For a summary of the tree health data assessed from the ground,
readers should refer to Dash et al. [12].
3.2. Spectral Indices
The spectral indices extracted from the control plots stayed within the ’no-change’ region for
both remote sensing data sources throughout the duration of the experiment (Figure 2). By contrast,
within the treated plots, spectral indices derived from both platforms exhibited a significant decline
over time. Indices derived from satellite imagery showed a marked reduction 75 days after treatment
that was well below the “no-change” region for NDVI and RENDVI and marginally below this region
for GNDVI. For UAV data all indices dropped below the no-change region 29 days after treatment.
This decline continued throughout the duration of the experiment for NDVI but levelled out 52 days
after treatment for the other two indices (Figure 2).
Analysing the changes in spectral indices within the canopies of tree clusters of different sizes
provides insight into the sensitivity of each sensor to physiological stress expression at different spatial
scales. Spectral indices derived from the UAV sensor showed the most marked changes over time
in NDVI. For all treatment cluster sizes, there was a marked decline in NDVI that moved below the
no-change region 29 days after treatment and continued to decline until the end of the experiment
(Figure 3). Values of GNDVI and RENDVI dropped below the no-change region at day 29. However,
after an initial decline, values of these indices stabilised after day 53 for most cluster sizes and exhibited
a sharp increase for the single tree clusters 64 days after treatment. The magnitude of reductions in
all indices scaled positively with cluster size. By day 87, reductions in GNDVI, NDVI, and RENDVI
were respectively, 0.89, 8.66, and 2.41 for single tree clusters and 3.18, 11.93, and 5.28 for clusters with
16 trees. Spectral indices derived from the satellite imagery were not sufficiently sensitive to detect
treatment changes in smaller clusters of trees. All indices remained within the ’no-change’ zone until
13 days after treatment. For cluster sizes of four or more trees, values for all indices were below the
’no-change’ zone for the next imagery acquisition 75 days after treatment. In contrast, index values
for clusters of two or fewer trees showed very little change 75 days after treatment. For these smaller
clusters, there was only marginal movement below the ’no-change’ zone for RENDVI in single tree
clusters (Figure 3). All indices showed greater reductions as cluster size increased and changes were
very muted for the two smallest clusters. Reductions in GNDVI, NDVI, and RENDVI 75 days after
treatment ranged from 0.29, 0.38, and 0.53 for single tree clusters to 2.01, 5.84, and 5.40 for clusters
with 16 trees. For both sensors and for all indices no decline below the ’no-change’ zone was observed
for the control plots where no herbicide was applied.
3.3. Spatial Resolution
Through resampling the UAV imagery to the spatial resolution of the RapidEye data we examined
the influence of spatial resolution on sensitivity to physiological stress onset. The general trends in
the spectral indices were very similar between the original UAV imagery and the imagery that was
resampled to 5 m (Figure 4). Restricting the analyses to data from 64 and 87 days after treatment
(as this spanned the time of the satellite imagery acquisition) the mean reduction in NDVI andRemote Sens. 2018, 10, 1216 10 of 22
RENDVI was greater in the resampled imagery than the original imagery for clusters with sixteen
(NDVI 14.6 vs. 11.9; RENDVI 6.39 vs. 5.28) and eight trees (NDVI 13.0 vs. 12.2; RENDVI 5.05 vs. 4.81).
In contrast, there was a smaller shift in these indices for resampled imagery with clusters of four
(NDVI 9.55 vs. 9.87; RENDVI 3.80 vs. 4.03), two (NDVI 8.18 vs. 9.63; RENDVI 3.18 vs. 3.86), and single
trees (NDVI 5.32 vs. 8.66; RENDVI 1.25 vs. 2.41). Similar patterns were observed for the less sensitive
metric GNDVI. Compared to the original imagery, the resampled imagery showed greater shifts
in GNDVI for clusters of eight or more trees and smaller shifts in GNDVI for clusters with four or
fewer trees.
Figure 2. Box and whisker plots showing the distribution of three de-trended spectral indices derived
from the two different sensors for control (n = 5) and treated (n = 25) plots over the course of the
experiment.The dashed lines show the lower bound of the ’no-change’ zone calculated from both
treated and control plots during the first time step. Change is deemed to have occurred once the upper
hinge of the box plot is below the ’no-change’ bound.
If the differences in the spectral indices observed between the satellite and UAV sensors were
solely the result of spatial resolution then we would expect to see very similar results extracted from
the satellite data and the resampled UAV imagery. In contrast to this expectation, resampling increased
the magnitude of these differences for the larger cluster sizes with eight and sixteen trees, as described
above. Consequently, the resampling process resulted in greater differences in all indices between
data derived from UAV and satellite sensors for these two cluster sizes. Although resampling the
UAV data decreased the magnitude of reductions in indices derived from smaller cluster sizes these
values were still substantially larger than those derived from satellite imagery for cluster sizes of four
trees (NDVI 9.55 vs. 2.12; RENDVI 3.80 vs. 2.57; GNDVI 2.30 vs. 0.56), two trees (NDVI 8.18 vs. 0.16;
RENDVI 3.18 vs. 1.82; GNDVI 0.9 vs. 0.14), and single trees (NDVI 5.32 vs. 0.38; RENDVI 1.25 vs. 0.53;
GNDVI 0.39 vs. 0.29).Remote Sens. 2018, 10, 1216 11 of 22
Figure 3. Each datum represents the 75th percentile of the spectral indices for an experimental cluster
size for both sensors across the duration of the study. The control plots where no trees were treated are
represented by the cluster size = 0. The red dashed line shows the 25th percentile of the pre-treatment
values and defines the lower bound of the ‘no-change’ region for a spectral index. The UAV results
plotted here have been previously displayed in [12].
Figure 4. Each tile displays the mean de-trended spectral indices calculated from a given cluster size
from a specific image source on a specific day. UAV-Resampled imagery shows the results extracted
from UAV imagery resampled to the spatial resolution of the RapidEye data.Remote Sens. 2018, 10, 1216 12 of 22
3.4. Comparing UAV and Satellite Imagery
Changes in de-trended vegetation indices were compared following imputation of missing values
in the experimental dataset (Figure 5). The values extracted from the UAV and satellite imagery were
very similar up to 20 days following treatment for all three spectral indices examined. Although
indices followed a similar general trend after this point, reductions in indices were greater for the data
from the UAV than the satellite sensor. This divergence in indices was most pronounced for NDVI.
Figure 5. The change in the de-trended vegetation indices and field measured discolouration plotted
against days after treatment. The solid line shows the mean at each time step and the shaded area
shows the confidence interval. For each spectral index the values relate to the mean value of pixels
within a cluster. For the ground measured discolouration the values represent observations of the
field technician.Remote Sens. 2018, 10, 1216 13 of 22
The relationship between satellite and UAV sensors was investigated using the imputed dataset
(Figure 6) where missing observations had been imputed using missForest. For both NDVI and
GNDVI, linear models described the relationship most accurately. For RENDVI the best model was a
second order polynomial. All developed models had high precision and model strength was greatest
for RENDVI (R2 = 0.92) followed by GNDVI (R2 = 0.84), and then NDVI (R2 = 0.82).
Figure 6. The relationship between UAV and satellite-derived vegetation indices. Each datum
represents the mean results extracted from a single time step for a tree cluster and the shaded area
shows the confidence interval.
A RF regression model was used to investigate the ability of both data sources to predict field
measured discolouration as reflected by visible discolouration. A single RF model (RMSE = 15.1,
bias = −0.02) was used to examine the relationship and this was able to account for 82.3% of theRemote Sens. 2018, 10, 1216 14 of 22
variability in the response. The conditional importance scores revealed that the NDVI index collected
from the UAV imagery was the most important variable for predicting discolouration (Figure 7).
The second most important predictor was the satellite RENDVI index, while the subsequent metrics
from both sensors were less important. RF regression models were also fitted using subsets of predictors
from only one data source to compare with the initial model fitted with all predictors. The model
using only UAV data was found to be more precise than the model that used spectral indices from the
satellite data (R2 = 0.84, RMSE = 15.04, bias = 0.13 vs. R2 = 0.73, RMSE = 18.56, bias = −0.03).
Figure 7. Conditional importance scores for all variables in the RF regression model.
4. Discussion
Spectral indices calculated from multispectral imagery were sufficiently robust to detect the onset
of physiological stress within this experiment. Conifer foliage may be damaged and change colour
due to a variety of agents, such as insects, root rot, fungi, and drought [83]. Independent of the causal
agent, under physiological stress foliar moisture declines, chlorophyll and other pigment molecules
break down, and this is followed by a break down of intra-cellular and cellular structures [63,83].
Hyperspectral imagery has been most often used to detect natural variation in pigments [84,85] and
elemental concentrations [86] although recent research suggests that precise models can be developed
for chlorophyll concentration within tree canopies from multispectral imagery [87]. Detection of the
more pronounced changes in needle chemistry and appearance associated with needle mortality can
be robustly predicted using multispectral imagery [19,88–92] which is consistent with our findings.
Results from this study, and previous analyses [12], showed NDVI and RENDVI to be the most
sensitive indicators of the onset of physiological stress. High values of NDVI in healthy foliage occur
due to the contrast of low leaf reflectance in the red band and higher reflectance within the near-infrared
band. As the foliage senesces, chlorophyll declines to cause an increase in red band reflectance resulting
in reductions in NDVI [46]. NDVI is one of the most widely-used spectral indices developed over
the last 40 years [93] and has been widely used to discriminate healthy from stressed foliage within
forests [94–98]. A large body of research has also demonstrated that increased reflectance in stressed
vegetation relative to healthy vegetation is also strongly evident along the red edge (680–750 nm)
region [61,88] and many studies have shown the red edge to be a more sensitive indicator of the onset
of stress than either the green or red bands [51]. Consequently, indices based on the red edge band haveRemote Sens. 2018, 10, 1216 15 of 22
also been widely used to detect stress [51] and defoliation resulting from insect attack and drought in
conifers [19,89,90,92].
Differences in the relative importance of indices were observed between platforms in our study.
NDVI derived from the UAV data was the most important predictor of field measured discolouration
while RENDVI was the strongest predictor of discolouration for the satellite data. These variations
may reflect the different spectral ranges covered by the two sensors examined. The greater importance
of RENDVI for the satellite data may be attributable to broader coverage of the red edge from this
platform (690–730 nm), compared to the red edge region covered by the UAV mounted sensor which is
limited to a narrower spectral window (707–727 nm). Previous studies have indicated that across a
range of higher plants the optimum wavelength for early detection of stress was 700 nm [99]. However,
an investigation of the stress detection in the conifer Balsam fir (Abies balsamea (L.) Mill) indicated that
reflectance at 711 nm was the most responsive [100]. These studies indicate that there may be some
variance in the precise wavelength where stress is most evident. As the RapidEye sensor covers the
entirety of the red edge region for the typical spectral response of plants it is likely to be highly sensitive
to changes in this region. The considerably narrower bands of the multispectral sensor could result
in a failure to detect very early spectral changes associated with physiological stress. Furthermore,
it should be noted that the incomplete time series provided by both sensors means that the sensitivity
of both platforms to the earliest signs of physiological stress could be confounded.
Although sensors on both platforms were sensitive to the symptoms of physiological stress,
changes in NDVI associated with stress were more pronounced using UAV data. By restricting analysis
to clusters larger than the pixel size of satellite imagery (4–16 trees) and to data from 64–87 days
after treatment the mean reduction across treated groups for imagery from the UAV, resampled UAV
imagery, and satellite imagery averaged −11.33, −12.67, and −3.94, respectively, for NDVI and −4.71,
−5.08, −3.84, respectively, for RENDVI. This result suggests that the lower sensitivity of the satellite
imagery was not solely attributable to differences in spatial resolution. Substantial differences remained
between the UAV and satellite data after the UAV data were resampled to the same spatial resolution.
Resampling minimised the effect of spatial resolution on sensitivity but the effects of differences caused
by atmospheric conditions and sensor spectral properties remained. It is logical to conclude that the
greater sensitivity of the UAV data to minor changes is attributable to the narrower bandwidths of the
UAV sensor compared to the RapidEye sensor and the proximity of the UAV platform to the subject
trees. However, the increase in the sensitivity of resampled spectral indices may be a product of the
resampling process and should not be interpreted as direct evidence that coarser spatial resolution
imagery has greater utility for stress monitoring. Alternatively, this could be the result of a greater
signal-to-noise ratio in coarser resolution imagery when the subject phenomenon is equal to or larger
than the pixel size of the imagery used. Further research should be undertaken to more precisely
partition the influences of spectral and spatial resolution on detection of stress so that the limits of
multispectral imagery can be more robustly defined.
Imagery acquired from a UAV was sensitive to stress in tree clusters of all sizes. It was interesting
to note that changes in NDVI were of a similar magnitude for cluster sizes greater than 25 m2 when
the imagery was resampled to a 5 m resolution. This finding is consistent with previous research
that demonstrated the utility of moderate resolution multispectral imagery for detection of stress.
Studies that have used multispectral imagery to detect insect damage suggest the optimal resolution to
be 2.4 m [48] for Pinus contorta dominated stands, 1.25–2.25 m for Eucalyptus spp. [98] and 1.75–2.3 m
for P. radiata [101]. However, the magnitude of the change in spectral indices for smaller clusters and
individual trees was lower for the resampled imagery.
Our findings clearly show that spectral indices derived from satellite imagery were only useful
for detecting stress in clusters of four or more trees. Satellite imagery has been successfully used to
detect major outbreaks of devastating insect damage that result in tree mortality. For example, recent
studies, that investigate mortality from outbreaks of mountain pine beetle Dendroctonus ponderosae
(Hopkins), showed classification accuracies ranging from 67–78% for medium resolution [91,102,103]Remote Sens. 2018, 10, 1216 16 of 22
to accuracies of 71–93% when high-resolution satellite imagery was used [20,21,90,104]. However, no
research has examined the precision of satellite imagery in detecting different sized clusters of stressed
or diseased trees under controlled conditions. Our results demonstrate the sensitivity of satellite
imagery for detecting stress symptoms and define the detection limits of stress in mature P. radiata for
this sensor type.
The success of forest surveillance is largely governed by being able to match the temporal, spatial,
and spectral characteristics of symptoms with an appropriate platform, sensor, and analytical method.
Results from this study demonstrate the complementary role that UAV and satellite platforms could
play in the detection of stress as symptoms develop. Solutions that can automate disease detection
from fine to medium resolution satellite imagery could be used to identify the location of potential
outbreaks. Once an outbreak has been detected, fine-scale imagery could be collected from UAVs to
regularly monitor an infection front across an area of interest in a cost-effective way. Data collection
from UAVs is considerably cheaper and can be collected in a greater range of weather conditions
than satellite data. Our research shows that the multispectral sensors studied are compatible and
can be used in a complementary manner. As a result, the UAV data can be used to augment the data
collected from the satellite platforms and enhance our understanding of the spatial distribution of
disease expression across forest ecosystems.
Our findings also show that data from the RapidEye satellite can be used to expand the range
of physiological stress monitoring that can be achieved using a UAV sensor. However, the RapidEye
sensor offers lower sensitivity as it was only suitable for detecting stress in larger clusters. When used in
combination these two technologies offer a solution for both widespread and fine-grained monitoring
of plantation forests. This system can be deployed to track physiological stress and to put in place
management activities to mitigate stress and increase forest productivity.
The knowledge developed through this experiment can be useful for the designers of multi-scale
detection systems for tracking biosecurity incursions. Intensive, high-resolution UAV surveys could be
deployed over high-risk sites and larger, less sensitive, satellite methods could be used across wider
areas. In combination, an approach of this type could provide an extremely practical tool as both an
early warning system and in monitoring the spread of an incursion.
5. Conclusions
Our research clearly shows that both UAV and satellite imagery have significant capability to
detect plant stress in mature P. radiata trees under controlled experimental conditions. This is evidenced
by the deviation of spectral indices from both sensors from the “no-change” region and from close
correlations with the field observations of tree health. This result suggests that a more comprehensive
monitoring platform can be developed by combining both sensor technologies. This would aid
precision forest management by enabling both coarse, broad-scale monitoring of disease progression
and fine-scale changes that might indicate new or early disease outbreaks. We found that smaller
scale instances of herbicide induced stress require finer spatial resolution imagery for detection.
This supports our hypothesis that the higher resolution UAV imagery provides greater sensitivity to
physiological stress than coarser satellite imagery. We have also defined the lower spatial bound for
physiological stress detection for the RapidEye satellite. In our experiment, the lower spatial bound
for stress detection was found to be the area represented by 4 or more trees, this equated to an area
of approximately 77 m2 on average. This finding suggests that, using this approach and the indices
examined, the RapidEye sensor is sensitive to changes at a spatial resolution that is approximately
twice the pixel size of the derived imagery. This may be useful for managers engaged in planning and
acquisition of satellite imagery. Our experiment showed that the UAV and satellite imagery data used
are complementary and can be used in combination to provide a more complete means of detecting
and monitoring physiological stress patterns in forested ecosystems. This multi-sensor approach will
lead to the augmentation and enhancement of current approaches based on a single platform.Remote Sens. 2018, 10, 1216 17 of 22
Author Contributions: Conceptualization, J.P.D. and M.S.W.; Data Curation, J.P.D.; Formal Analysis, J.P.D.
and G.D.P.; Investigation, J.P.D., G.D.P. and M.S.W.; Methodology, J.P.D.; Visualization, J.P.D.;
Writing—Original Draft, J.P.D.; Writing—Review & Editing, G.D.P. and M.S.W.
Funding: This study was funded as part of the Growing Confidence in Forestry’s Future (GCFF) programme
(Contract number: CO4X1306). GCFF was co-funded by the Forest Growers Levy Trust and the New Zealand
Ministry of Business, Innovation and Employment (MBIE).
Acknowledgments: The authors recognise the contributions of Toby Stovold, Kane Fleet, Marie Heaphy, Mark
Miller and Rod Brownlie who completed all field work and Heidi Dungey who helped develop the study methods.
Scion colleague Stefan Gous is acknowledged for offering advice on herbicide application rates. We are grateful to
the forest managers for providing access and allowing us to complete this study in their forest.
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Moore, J.R.; Dash, J.P.; Lee, J.R.; McKinley, R.B.; Dungey, H.S. Quantifying the influence of seedlot and stand
density on growth, wood properties and the economics of growing radiata pine. Forestry 2017, 91, 327–340,
doi:10.1093/forestry/cpx016. [CrossRef]
2. Watt, M.S.; Kimberley, M.O.; Dash, J.P.; Harrison, D. Spatial prediction of optimal final stand density
for even-aged plantation forests using productivity indices. Can. J. For. Res. 2017, 47, 527–535,
doi:10.1139/cjfr-2016-0220. [CrossRef]
3. Klapste, J.; Suontama, M.; Telfer, E.; Graham, N.; Low, C.; Stovold, T.; McKinley, R.; Dungey, H. Exploration of
genetic architecture through sib-ship reconstruction in advanced breeding population of Eucalyptus nitens.
PLoS ONE 2017, 12, e0185137, doi:10.1371/journal.pone.0185137. [CrossRef] [PubMed]
4. Li, Y.; Xue, J.; Clinton, P.W.; Dungey, H.S. Genetic parameters and clone by environment interactions for
growth and foliar nutrient concentrations in radiata pine on 14 widely diverse New Zealand sites. Tree Genet.
Genomes 2015, 11, 10, doi:10.1007/s11295-014-0830-1. [CrossRef]
5. Maggard, A.O.; Will, R.E.; Wilson, D.S.; Meek, C.R.; Vogel, J.G. Fertilization can compensate for decreased
water availability by increasing the efficiency of stem volume production per unit of leaf area for loblolly
pine (Pinus taeda) stands. Can. J. For. Res. 2017, 47, 445–457, doi:10.1139/cjfr-2016-0422. [CrossRef]
6. Watt, M.S.; Dash, J.P.; Bhandari, S.; Watt, P. Comparing parametric and non-parametric methods of predicting
Site Index for radiata pine using combinations of data derived from environmental surfaces, satellite imagery
and airborne laser scanning. For. Ecol. Manag. 2015, 357, 1–9, doi:10.1016/j.foreco.2015.08.001. [CrossRef]
7. Watt, M.S.; Dash, J.P.; Watt, P.; Bhandari, S. Multi-sensor modelling of a forest productivity index for radiata
pine plantations. N. Z. J. For. Sci. 2016, 46, 9, doi:10.1186/s40490-016-0065-z. [CrossRef]
8. Pearse, G.D.; Morgenroth, J.; Watt, M.S.; Dash, J.P. Optimising prediction of forest leaf area index from
discrete airborne lidar. Remote Sens. Environ. 2017, 200, 220–239, doi:10.1016/j.rse.2017.08.002. [CrossRef]
9. Paoletti, E. Impact of ozone on Mediterranean forests: A review. Environ. Pollut. 2006, 144, 463–474,
doi:10.1016/j.envpol.2005.12.051. [CrossRef] [PubMed]
10. Marco, A.D.; Vitale, M.; Popa, I.; Anav, A.; Badea, O.; Silaghi, D.; Leca, S.; Screpanti, A.; Paoletti, E.
Ozone exposure affects tree defoliation in a continental climate. Sci. Total Environ. 2017, 596–597, 396–404,
doi:10.1016/j.scitotenv.2017.03.135. [CrossRef] [PubMed]
11. Moore, J.R.; Watt, M.S. Modelling the influence of predicted future climate change on the risk of wind
damage within New Zealand’s planted forests. Glob. Chang. Biol. 2015, 21, 3021–3035, doi:10.1111/gcb.12900.
[CrossRef] [PubMed]
12. Dash, J.P.; Watt, M.S.; Pearse, G.D.; Heaphy, M.; Dungey, H.S. Assessing very high resolution UAV imagery
for monitoring forest health during a simulated disease outbreak. ISPRS J. Photogramm. Remote Sens. 2017,
131, 1–14, doi:10.1016/j.isprsjprs.2017.07.007. [CrossRef]
13. Bulman, L.S.; Bradshaw, R.E.; Fraser, S.; Martín-García, J.; Barnes, I.; Musolin, D.L.; La Porta, N.; Woods, A.J.;
Diez, J.J.; Koltay, A.; et al. A worldwide perspective on the management and control of Dothistroma
needle blight. For. Pathol. 2016, 46, 472–488, doi:10.1111/efp.12305. [CrossRef]
14. Coops, N.; Stanford, M.; Old, K.; Dudzinski, M.; Culvenor, D.; Stone, C. Assessment of Dothistroma
Needle Blight of Pinus radiata Using Airborne Hyperspectral Imagery. Phytopathology 2003, 93, 1524–1532,
doi:10.1094/PHYTO.2003.93.12.1524. [CrossRef] [PubMed]You can also read
